Digital Watermarking Scheme Based on Machine Learning for the IHC Evaluation Criteria

Abstract

Digital watermarking is a technique used for embedding information in digital content and protecting its copyright. The important issues to be considered are robustness, quality and capacity. Our goal is to satisfy these requirements according to the Information Hiding and its Criteria for evaluation (IHC) criteria. In this study, we evaluate our watermarking scheme along the IHC criteria Ver.3 as the primary step. Although image watermarking techniques based on machine learning already exist, their robustness against desynchronization attacks such as cropping, rotation, and scaling is still one of the most challenging issues. We propose a watermarking scheme based on machine learning which also has cropping tolerance. First, the luminance space of the image is decomposed by one level through wavelet transform. Then, a bit of the watermark and the marker for synchronization are embedded or extracted by adjusting or comparing the relation between the embedded coefficients value of the LL space and the output coefficients value of the trained machine learning model. This model can well memorize the relationship between its selected coefficients and the neighboring coefficients. The marker for synchronization is embedded in a latticed format in the LL space. Binarization processing is performed on the watermarked image to find the lattice-shaped marker and synchronize it against cropping. Our experimental results showed that there were no errors in 10HDTV-size areas after the second decompression.